Prime Minister Benjamin Netanyahu and Brig.-Gen. (res.) Erez Eskel, head of the National AI Directorate, launched Israel's National AI Program on August 9, 2026, in a joint ceremony that framed the state's role in the AI era not as a rulemaker but as a builder. "AI is not just another technology; it is going to change the world," Netanyahu said, comparing the shift to the Industrial Revolution (Jerusalem Post). The program, which the government had formally approved via Government Decision 4255 on June 16, 2026, commits Israel to roughly 100,000 GPUs for sovereign AI data centers and a new university AI degree track launching in October 2026 aimed at reskilling a workforce the government estimates could see 1 to 4 million Israelis affected by AI adoption (Let's Data Science; JNS).
The Case for the Builder-State Model
There's a real argument for what Israel is doing here, and it deserves to be stated fairly before it's critiqued. The EU's AI Act spent years litigating risk tiers and conformity assessments before a single euro went toward sovereign compute; Israel, by contrast, is trying to compress cyber-era muscle memory — public-private co-investment, defense-linked R&D, fast state procurement — into an AI buildout. Eskel's framing at the launch was telling: "the human capital in the State of Israel...is our true engine" (Israel National News), and Netanyahu explicitly benchmarked the plan against Israel's cybersecurity success, a sector where the state built capacity and let a light regulatory touch (three narrow sandboxes — autonomous vehicles, fintech, and general use — rather than a horizontal AI statute) do the rest. Building first and regulating narrowly, where needed, avoids the EU's now-visible trap: a compliance regime operational years before the compute or talent existed to make it matter.
Where the Model Meets a Hard Constraint
But a state that decides to become an infrastructure builder inherits a builder's problems, and Israel's is showing up faster than the GPUs. Weeks after the government approved its AI acceleration plan, Israel's Electricity Authority disclosed that pending data-center connection requests had reached roughly 27,000 megawatts — three times the country's current average consumption of about 9,000 MW — against 1.5 gigawatts of capacity actually committed through 2035. The regulator's response was a 140-day freeze on processing new data-center connection requests above 8 MW, running from July 2026 into early December, while it works out an allocation policy. "These are insane numbers, and there is no scenario in which we can catch up with such demand," one senior electricity official said; a data-center industry figure called the freeze "worse than receiving a negative answer," since open uncertainty is harder to plan around than a denial (Calcalist/CTech). That freeze is not incidental to the AI program — it is the National AI Program's own resource constraint made visible, and it is a regulatory act, full stop: a state agency deciding who gets to build and when.
The chip side of the equation cuts the other way. In December 2025, Israel was admitted to the nine-country "Pax Silica" coalition — alongside the US, Japan, South Korea, the Netherlands, Singapore, the UK, Australia, and the UAE — moving it off the Biden-era "yellow list" of countries facing case-by-case export licensing and into what officials privately call "gold club" status: largely unrestricted flow of advanced semiconductors and AI infrastructure among members (Calcalist/CTech). That access removes one bottleneck Israel cannot control domestically. It does not remove the other one, which is entirely domestic: grid capacity, permitting speed, and who the state lets build first.
The Real Risk Isn't a Missing AI Law
Outside estimates put the cost of the 100,000-GPU target at $20-30 billion or more — Tel Aviv University's Prof. Nadav Cohen called it "a very ambitious goal that could easily cost tens of billions," and cautioned that chasing full independence "across all layers" of the AI stack is "overly ambitious" for any single country, urging Israel to concentrate on areas — physical AI, edge systems, defense-linked research — where it already leads rather than trying to replicate the entire US stack at home (Ynet). That is the more useful caution than anything an EU-style AI Act would offer. Israel does not need a horizontal AI statute to make this program work; its sandbox approach to sector-specific AI risk remains proportionate to where actual harms have shown up so far. What it needs is the same rigor applied to the state's own role as an infrastructure allocator that it is asking of the private sector: transparent, published criteria for who gets grid capacity and when, rather than a blanket freeze followed by ad hoc exceptions — because a 140-day silence on power allocation is a worse signal to investors than any compliance checklist would be.
The National AI Program is the right instinct: treat AI capacity as an infrastructure problem the state can accelerate, not primarily a risk to pre-authorize. But Netanyahu's government has, in effect, appointed itself referee of the scarcest input in its own plan. Whether Israel becomes the AI power it wants to be will be decided less by Knesset debate over an AI law than by whether the Electricity Authority's allocation policy — due by early December 2026 — is transparent enough that the 100,000-GPU commitment survives contact with the grid.